A novel data-driven sample augmentation method using interpretable space and adversarial mechanism for dioxin risk warning

Dioxin risk warning plays a crucial role in ensuring the long-term sustainability of municipal solid waste incineration (MSWI) plants. However, dioxin, which is difficult to detect in real time, poses significant challenges in developing data-driven models. One widely employed solution to overcoming the limitations of data-driven modeling is the virtual sample generation (VSG) method. Nevertheless, incomplete generative procedures and unreliable randomness produce difficulties for researchers utilizing VSG. Addressing these issues, this article proposes a novel data augmentation method based on interpretable space and adversarial mechanism. Initially, two-dimensional spatial latent samples are derived using a variational autoencoder to aid visualization. Subsequently, two-dimensional virtual latent samples are generated via multi-angle rotations in diverse circular spaces to enhance interpretability. Next, these two-dimensional virtual latent samples are input into the generative adversarial network to produce candidate virtual samples in their original dimensions. Finally, co-training is employed to meticulously select high-quality candidate virtual samples. Experimental verification of the proposed VSG method utilizes real dioxin datasets of the MSWI process, demonstrating its effectiveness.